Winner of the

DeepBreath—automated detection of respiratory pathology

Alain Gervaix et Al.

Summary

This high impact medical research article describes and analyzes the design of an innovative deep learning model designed to identify lung sounds in children. As traditional methods of understanding these sounds have been subjective and laden with unclear terms, DeepBreath brings clarity in identifying the most common lung-related diseases – pneumonia, bronchiolitis, and wheezing disorders.

The machine was trained based on 35.9 hours of audio recordings from 572 young patients. It employs a neural network and logistic regression and delves into recordings from eight different chest locations, discerning patterns and makes predictions at the individual patient level.

The patient pool consisted of 29% healthy individuals and 71% individuals with respiratory issues coming fromSwitzerland and Brazil, whose processed results were validated on patients from Senegal, Cameroon, and Morocco.

DeepBreath showcased an impressive ability to distinguish between healthy and problematic breathing and the results were promising in identifying specific issues.

Delving deeper into the analysis, DeepBreath appeared to also be able to capture meaningful physiological information.

In essence, DeepBreath is a user-friendly, deep learning framework that effectively identifies audio signatures indicative of respiratory issues in children, and works as a model for further implementation of deep learning machines in scarce resource contexts where medical knowledge is needed most and not always available.

Other Winners

Previous
Next